meshllm/moe-rankings

MoE Rankings

meshllm/moe-rankings is a public dataset of derived Mixture-of-Experts routing metadata for published model artifacts.

The dataset stores ranking artifacts produced by llama-moe-analyze so tools such as mesh-llm can discover expert-hotness rankings for exact model revisions without recomputing them locally.

Purpose

This dataset exists to provide:

• immutable MoE expert rankings keyed by exact source model revision
• a canonical archive of published ranking artifacts
• reusable metadata for routing, sharding, and MoE placement experiments

The dataset is not a model mirror and does not store original model weights.

Identity Model

Each artifact is identified by:

• source_repo
• source_revision
• format
• distribution_id
• analyzer_id

For GGUF models, distribution_id is the normalized model distribution name, usually the GGUF filename stem with any shard suffix removed.

Layout

Artifacts are stored under:

data/<source_namespace>/<source_repo_name>/<source_revision>/<format>/<distribution_id>/<analyzer_id>/

Each artifact directory contains:

• metadata.json
• ranking.csv
• run.log

Example:

data/Flexan/kshitijthakkar-qwen3.5-moe-0.87B-d0.8B-GGUF/a9b8adbec2cc87479c772dac1944f313b4036c26/gguf/qwen3.5-moe-0.87B-d0.8B.Q2_K/micro-v1/

Artifact Semantics

ranking.csv

Normalized expert ranking output with columns:

expert_id,total_mass,mass_fraction,selection_count

Sorted by hottest experts first.

metadata.json

Validation and provenance metadata, including:

• exact source repo and commit
• analyzed distribution id
• file list and hashes
• analyzer id
• prompt set id
• token count
• local analyzer source details

run.log

Raw execution log for debugging and auditing.

Analyzer Policy

Current canonical analyzer:

• micro-v1

micro-v1 is tied to a fixed built-in prompt set and should be comparable across runs. Any meaningful change to prompts or semantics should produce a new analyzer id such as micro-v2.

Immutability

Artifacts in this dataset are intended to be immutable.

• A new source model commit uses a new source_revision path.
• A new analysis method or incompatible prompt set uses a new analyzer_id.
• Existing published artifacts should not be overwritten with different content.

Intended Consumers

• mesh-llm
• MoE sharding and routing tools
• benchmarking and evaluation pipelines
• researchers comparing expert distributions across quantizations and revisions

Notes

• This dataset stores derived metadata, not original model weights.
• Some logs may be verbose because they preserve upstream tool output for reproducibility.
• Model-repo colocated sidecars may exist separately, but this dataset is the canonical system of record.